On Feb 27, 2013, at 3:19 AM CST, Pavan Balaji wrote:

On 02/25/2013 08:41 AM US Central Time, Jed Brown wrote:

...

It sounds like you are binding a process. Suppose hydra binds my process to cpuset 0xf and my process pthread_creates a total of three threads. If I bind those threads to 0x1, 0x2, 0x4, we still have the MPICH_ASYNC_PROGRESS running with affinity 0xf, and as far as I know, there is no way to bind it to 0x8.

It would be interesting to be able to specify this, both to place the progress threads close to NICs and to make sure that progress threads do not disrupt application threads (to reduce jitter and possibly improve performance).

There's no plan to support per-thread binding. This requires mpich to keep track of thread IDs, which AFAICT, we don't, unless I'm forgetting some piece deep inside MPICH (Dave or Jim should correct me, if I'm wrong).

We do use thread-ids in some cases, but we definitely aren't tracking the overall set of threads in existence or otherwise manipulating them. I think we mainly use thread-ids to deal with thread-local storage in some cases. -Dave